On Wed, Dec 7, 2016, at 18:32, Eric Dumazet wrote:
> On Wed, 2016-12-07 at 17:09 +0000, David Laight wrote:
> > From: Paolo Abeni
> > > Sent: 06 December 2016 17:08
> > ...
> > > @@ -79,6 +82,9 @@ struct udp_sock {
> > > int (*gro_complete)(struct sock *sk,
> > > struct sk_buff *skb,
> > > int nhoff);
> > > +
> > > + /* since we are prone to drops, avoid dirtying any sk cacheline */
> > > + atomic_t drops ____cacheline_aligned_in_smp;
> > > };
> >
> > Isn't that likely to create a large hole on systems with large cache lines.
> > (Same as any other use of ____cacheline_aligned_in_smp.)
>
> Yes, I would like to avoid that, unless we come to the conclusion it is
> absolutely needed.
>
> I feel that we could simply use a pointer, and allocate memory on
> demand, since many sockets do not ever experience a drop.
>
> The pointer could stay in a read mostly section.
>
> We even could use per cpu or node counter for some heavy drop cases.
I had the same idea while discussing that with Paolo, merely using an
*atomic_t = kmalloc(sizeof(atomic_t)) out of band of the socket.
My fear was that those could be aggregated by the slab cache into one
cache line, causing even more heating on cachelines.
Bye,
Hannes
